{
  "run_id": "20260609_185647",
  "version": "v4",
  "base_model": "openai/gpt-oss-20b",
  "strategy": "LoRA on lm_head ONLY \u2014 backbone frozen \u2014 BF16",
  "key_fixes_vs_v3_fixed": [
    "lm_head weight untied (detach+clone) BEFORE get_peft_model()",
    "resize_token_embeddings tie-re-introduction guard added",
    "lora_r raised 16\u219232, lora_alpha raised 64\u2192128",
    "grad_accum raised 16\u219232, lr raised 1e-4\u21922e-4",
    "max_grad_norm tightened 0.5\u21920.3, weight_decay=0.01 added",
    "EarlyStoppingCallback added",
    "scheduler changed cosine\u2192cosine_with_restarts",
    "optimizer changed adamw_torch\u2192adamw_torch_fused"
  ],
  "lora_r": 32,
  "lora_alpha": 128,
  "lora_dropout": 0.1,
  "target_modules": [
    "lm_head"
  ],
  "epochs": 15,
  "learning_rate": 0.0002,
  "max_grad_norm": 0.3,
  "weight_decay": 0.01,
  "batch_size": 1,
  "gradient_accumulation": 32,
  "effective_batch_size": 32,
  "max_seq_len": 2048,
  "train_samples": 5291,
  "val_samples": 5291,
  "test_samples": 5291,
  "train_metrics": {
    "train_runtime": 24397.0057,
    "train_samples_per_second": 3.253,
    "train_steps_per_second": 0.102,
    "total_flos": 2.486936885562409e+18,
    "train_loss": 28.171284981233526,
    "epoch": 15.0
  },
  "eval_metrics": {
    "eval_loss": 1.5277799367904663,
    "eval_runtime": 412.216,
    "eval_samples_per_second": 12.836,
    "eval_steps_per_second": 12.836,
    "epoch": 15.0
  },
  "test_results_summary": {
    "mean_test_loss": 1.593661,
    "exact_match_rate": 0.0,
    "mean_token_f1": 0.270723
  },
  "final_adapter_dir": "/mnt/DATA/saurabh/aditya/SpaceLLM/fine_tuning_v2/outputs/spacellm_lora_final",
  "checkpoints_dir": "/mnt/DATA/saurabh/aditya/SpaceLLM/fine_tuning_v2/outputs/checkpoints",
  "graphs_dir": "/mnt/DATA/saurabh/aditya/SpaceLLM/fine_tuning_v2/outputs/graphs",
  "log_file": "/mnt/DATA/saurabh/aditya/SpaceLLM/fine_tuning_v2/outputs/logs/train_20260609_185647.log"
}